Foreground removal is a challenge for 21‐cm tomography of the high‐redshift Universe. We use archival Giant Metrewave Radio Telescope (GMRT) data (obtained for completely different astronomical goals) to estimate the foregrounds at a redshift of ∼1. The statistic we use is the cross power spectrum between two frequencies separated by Δν at the angular multipole ℓ, or equivalently the multi‐frequency angular power spectrum Cℓ(Δν). An earlier measurement of Cℓ(Δν) using these data had revealed the presence of oscillatory patterns along Δν, which turned out to be a severe impediment for foreground removal. Using the same data, in this paper we show that it is possible to considerably reduce these oscillations by suppressing the sidelobe response of the primary antenna elements. The suppression works best at the angular multipoles ℓ for which there is a dense sampling of the u−v plane. For three angular multipoles ℓ= 1405, 1602 and 1876, this sidelobe suppression along with a low order polynomial fitting completely results in residuals of (≤ 0.02 mK2), consistent with the noise at the 3σ level. Since the polynomial fitting is done after estimation of the power spectrum it can be ensured that the estimation of the H i signal is not biased. The corresponding 99 per cent upper limit on the H i signal is , where is the mean neutral fraction and b is the bias.